#!/usr/bin/env python import os import random import gradio as gr import numpy as np import PIL.Image import torch import torchvision.transforms.functional as TF from diffusers import ( AutoencoderKL, EulerAncestralDiscreteScheduler, StableDiffusionXLAdapterPipeline, T2IAdapter, ) from PIL import Image from src.tryon_pipeline import StableDiffusionXLInpaintPipeline as TryonPipeline from src.unet_hacked_garmnet import UNet2DConditionModel as UNet2DConditionModel_ref from src.unet_hacked_tryon import UNet2DConditionModel from transformers import ( CLIPImageProcessor, CLIPVisionModelWithProjection, CLIPTextModel, CLIPTextModelWithProjection, ) from diffusers import DDPMScheduler,AutoencoderKL from typing import List import torch import os from transformers import AutoTokenizer import spaces import numpy as np from utils_mask import get_mask_location from torchvision import transforms import apply_net from preprocess.humanparsing.run_parsing import Parsing from preprocess.openpose.run_openpose import OpenPose from detectron2.data.detection_utils import convert_PIL_to_numpy,_apply_exif_orientation from torchvision.transforms.functional import to_pil_image base_path = 'yisol/IDM-VTON' example_path = os.path.join(os.path.dirname(__file__), 'example') def pil_to_binary_mask(pil_image, threshold=0): np_image = np.array(pil_image) grayscale_image = Image.fromarray(np_image).convert("L") binary_mask = np.array(grayscale_image) > threshold mask = np.zeros(binary_mask.shape, dtype=np.uint8) for i in range(binary_mask.shape[0]): for j in range(binary_mask.shape[1]): if binary_mask[i,j] == True : mask[i,j] = 1 mask = (mask*255).astype(np.uint8) output_mask = Image.fromarray(mask) return output_mask unet = UNet2DConditionModel.from_pretrained( base_path, subfolder="unet", torch_dtype=torch.float16, ) unet.requires_grad_(False) tokenizer_one = AutoTokenizer.from_pretrained( base_path, subfolder="tokenizer", revision=None, use_fast=False, ) tokenizer_two = AutoTokenizer.from_pretrained( base_path, subfolder="tokenizer_2", revision=None, use_fast=False, ) noise_scheduler = DDPMScheduler.from_pretrained(base_path, subfolder="scheduler") text_encoder_one = CLIPTextModel.from_pretrained( base_path, subfolder="text_encoder", torch_dtype=torch.float16, ) text_encoder_two = CLIPTextModelWithProjection.from_pretrained( base_path, subfolder="text_encoder_2", torch_dtype=torch.float16, ) image_encoder = CLIPVisionModelWithProjection.from_pretrained( base_path, subfolder="image_encoder", torch_dtype=torch.float16, ) vae = AutoencoderKL.from_pretrained(base_path, subfolder="vae", torch_dtype=torch.float16, ) # "stabilityai/stable-diffusion-xl-base-1.0", UNet_Encoder = UNet2DConditionModel_ref.from_pretrained( base_path, subfolder="unet_encoder", torch_dtype=torch.float16, ) parsing_model = Parsing(0) openpose_model = OpenPose(0) UNet_Encoder.requires_grad_(False) image_encoder.requires_grad_(False) vae.requires_grad_(False) unet.requires_grad_(False) text_encoder_one.requires_grad_(False) text_encoder_two.requires_grad_(False) tensor_transfrom = transforms.Compose( [ transforms.ToTensor(), transforms.Normalize([0.5], [0.5]), ] ) pipe = TryonPipeline.from_pretrained( base_path, unet=unet, vae=vae, feature_extractor= CLIPImageProcessor(), text_encoder = text_encoder_one, text_encoder_2 = text_encoder_two, tokenizer = tokenizer_one, tokenizer_2 = tokenizer_two, scheduler = noise_scheduler, image_encoder=image_encoder, torch_dtype=torch.float16, ) pipe.unet_encoder = UNet_Encoder @spaces.GPU def start_tryon(dict,garm_img,garment_des,is_checked,is_checked_crop,denoise_steps,seed,area): device = "cuda" openpose_model.preprocessor.body_estimation.model.to(device) pipe.to(device) pipe.unet_encoder.to(device) OUTPUT_WIDTH, OUTPUT_HEIGHT = dict.size garm_img= garm_img.convert("RGB").resize((768,1024)) human_img_orig = dict.convert("RGB").resize((768,1024)) # human_img_orig = dict["background"].convert("RGB") if is_checked_crop: width, height = human_img_orig.size target_width = int(min(width, height * (3 / 4))) target_height = int(min(height, width * (4 / 3))) left = (width - target_width) / 2 top = (height - target_height) / 2 right = (width + target_width) / 2 bottom = (height + target_height) / 2 cropped_img = human_img_orig.crop((left, top, right, bottom)) crop_size = cropped_img.size human_img = cropped_img.resize((768,1024)) else: human_img = human_img_orig.resize((768,1024)) if is_checked: keypoints = openpose_model(human_img.resize((384,512))) model_parse, _ = parsing_model(human_img.resize((384,512))) mask, mask_gray = get_mask_location('hd', area, model_parse, keypoints) mask = mask.resize((768,1024)) # else: # mask = pil_to_binary_mask(dict['layers'][0].convert("RGB").resize((768, 1024))) # mask = transforms.ToTensor()(mask) # mask = mask.unsqueeze(0) mask_gray = (1-transforms.ToTensor()(mask)) * tensor_transfrom(human_img) mask_gray = to_pil_image((mask_gray+1.0)/2.0) human_img_arg = _apply_exif_orientation(human_img.resize((384,512))) human_img_arg = convert_PIL_to_numpy(human_img_arg, format="BGR") args = apply_net.create_argument_parser().parse_args(('show', './configs/densepose_rcnn_R_50_FPN_s1x.yaml', './ckpt/densepose/model_final_162be9.pkl', 'dp_segm', '-v', '--opts', 'MODEL.DEVICE', 'cuda')) # verbosity = getattr(args, "verbosity", None) pose_img = args.func(args,human_img_arg) pose_img = pose_img[:,:,::-1] pose_img = Image.fromarray(pose_img).resize((768,1024)) with torch.no_grad(): # Extract the images with torch.cuda.amp.autocast(): with torch.no_grad(): prompt = "model is wearing " + garment_des negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality" with torch.inference_mode(): ( prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds, ) = pipe.encode_prompt( prompt, num_images_per_prompt=1, do_classifier_free_guidance=True, negative_prompt=negative_prompt, ) prompt = "a photo of " + garment_des negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality" if not isinstance(prompt, List): prompt = [prompt] * 1 if not isinstance(negative_prompt, List): negative_prompt = [negative_prompt] * 1 with torch.inference_mode(): ( prompt_embeds_c, _, _, _, ) = pipe.encode_prompt( prompt, num_images_per_prompt=1, do_classifier_free_guidance=False, negative_prompt=negative_prompt, ) pose_img = tensor_transfrom(pose_img).unsqueeze(0).to(device,torch.float16) garm_tensor = tensor_transfrom(garm_img).unsqueeze(0).to(device,torch.float16) generator = torch.Generator(device).manual_seed(seed) if seed is not None else None images = pipe( prompt_embeds=prompt_embeds.to(device,torch.float16), negative_prompt_embeds=negative_prompt_embeds.to(device,torch.float16), pooled_prompt_embeds=pooled_prompt_embeds.to(device,torch.float16), negative_pooled_prompt_embeds=negative_pooled_prompt_embeds.to(device,torch.float16), num_inference_steps=denoise_steps, generator=generator, strength = 1.0, pose_img = pose_img.to(device,torch.float16), text_embeds_cloth=prompt_embeds_c.to(device,torch.float16), cloth = garm_tensor.to(device,torch.float16), mask_image=mask, image=human_img, height=1024, width=768, ip_adapter_image = garm_img.resize((768,1024)), guidance_scale=2.0, )[0] if is_checked_crop: out_img = images[0].resize(crop_size) human_img_orig.paste(out_img, (int(left), int(top))) return human_img_orig.resize((OUTPUT_WIDTH, OUTPUT_HEIGHT)) else: return images[0].resize((OUTPUT_WIDTH, OUTPUT_HEIGHT)) # return images[0], mask_gray garm_list = os.listdir(os.path.join(example_path,"cloth")) garm_list_path = [os.path.join(example_path,"cloth",garm) for garm in garm_list] image_blocks = gr.Blocks(css="style.css").queue() with image_blocks as demo: gr.Markdown("## MyFit-AI") with gr.Row(): with gr.Column(): imgs = gr.Image(label='Human', type="pil") garm_img = gr.Image(label="Garment", type="pil") with gr.Row(elem_id="prompt-container"): with gr.Row(): prompt = gr.Textbox(placeholder="Description of garment ex) Neck T-shirts", show_label=False, elem_id="prompt") example = gr.Examples( inputs=garm_img, examples_per_page=8, examples=garm_list_path) try_button = gr.Button(value="Try-on", variant="primary") with gr.Accordion(label="Advanced Settings", open=False): with gr.Row(): denoise_steps = gr.Number(label="Denoising Steps", minimum=20, maximum=40, value=30, step=1) seed = gr.Number(label="Seed", minimum=-1, maximum=2147483647, step=1, value=42) with gr.Row(): area = gr.Dropdown(["upper_body","lower_body"], value="upper_body", label="garment zone") with gr.Row(): is_checked = gr.Checkbox(label="Yes", info="Use auto-generated mask (Takes 5 seconds)",value=True,visible=False) with gr.Row(): is_checked_crop = gr.Checkbox(label="Yes", info="Use auto-crop & resizing",value=True,visible=False) with gr.Column(): image_out = gr.Image(label="Output", elem_id="output-img",show_share_button=True) try_button.click(fn=start_tryon, inputs=[imgs, garm_img, prompt, is_checked,is_checked_crop, denoise_steps, seed, area], outputs=[image_out], api_name='tryon') # Doodly - T2I-Adapter-SDXL Sketch style_list = [ { "name": "(No style)", "prompt": "{prompt}", "negative_prompt": "", }, { "name": "Cinematic", "prompt": "cinematic still {prompt} . emotional, harmonious, vignette, highly detailed, high budget, bokeh, cinemascope, moody, epic, gorgeous, film grain, grainy", "negative_prompt": "anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured", }, { "name": "3D Model", "prompt": "professional 3d model {prompt} . octane render, highly detailed, volumetric, dramatic lighting", "negative_prompt": "ugly, deformed, noisy, low poly, blurry, painting", }, { "name": "Anime", "prompt": "anime artwork {prompt} . anime style, key visual, vibrant, studio anime, highly detailed", "negative_prompt": "photo, deformed, black and white, realism, disfigured, low contrast", }, { "name": "Digital Art", "prompt": "concept art {prompt} . digital artwork, illustrative, painterly, matte painting, highly detailed", "negative_prompt": "photo, photorealistic, realism, ugly", }, { "name": "Photographic", "prompt": "cinematic photo {prompt} . 35mm photograph, film, bokeh, professional, 4k, highly detailed", "negative_prompt": "drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly", }, { "name": "Pixel art", "prompt": "pixel-art {prompt} . low-res, blocky, pixel art style, 8-bit graphics", "negative_prompt": "sloppy, messy, blurry, noisy, highly detailed, ultra textured, photo, realistic", }, { "name": "Fantasy art", "prompt": "ethereal fantasy concept art of {prompt} . magnificent, celestial, ethereal, painterly, epic, majestic, magical, fantasy art, cover art, dreamy", "negative_prompt": "photographic, realistic, realism, 35mm film, dslr, cropped, frame, text, deformed, glitch, noise, noisy, off-center, deformed, cross-eyed, closed eyes, bad anatomy, ugly, disfigured, sloppy, duplicate, mutated, black and white", }, { "name": "Neonpunk", "prompt": "neonpunk style {prompt} . cyberpunk, vaporwave, neon, vibes, vibrant, stunningly beautiful, crisp, detailed, sleek, ultramodern, magenta highlights, dark purple shadows, high contrast, cinematic, ultra detailed, intricate, professional", "negative_prompt": "painting, drawing, illustration, glitch, deformed, mutated, cross-eyed, ugly, disfigured", }, { "name": "Manga", "prompt": "manga style {prompt} . vibrant, high-energy, detailed, iconic, Japanese comic style", "negative_prompt": "ugly, deformed, noisy, blurry, low contrast, realism, photorealistic, Western comic style", }, ] styles = {k["name"]: (k["prompt"], k["negative_prompt"]) for k in style_list} STYLE_NAMES = list(styles.keys()) DEFAULT_STYLE_NAME = "(No style)" def apply_style(style_name: str, positive: str, negative: str = "") -> tuple[str, str]: p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME]) return p.replace("{prompt}", positive), n + negative device = torch.device("cuda" if torch.cuda.is_available() else "cpu") if torch.cuda.is_available(): model_id = "stabilityai/stable-diffusion-xl-base-1.0" adapter = T2IAdapter.from_pretrained( "TencentARC/t2i-adapter-sketch-sdxl-1.0", torch_dtype=torch.float16, variant="fp16" ) scheduler = EulerAncestralDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler") pipe = StableDiffusionXLAdapterPipeline.from_pretrained( model_id, vae=AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16), adapter=adapter, scheduler=scheduler, torch_dtype=torch.float16, variant="fp16", ) pipe.to(device) else: pipe = None MAX_SEED = np.iinfo(np.int32).max def randomize_seed_fn(seed: int, randomize_seed: bool) -> int: if randomize_seed: seed = random.randint(0, MAX_SEED) return seed def run( image: PIL.Image.Image, prompt: str, negative_prompt: str, style_name: str = DEFAULT_STYLE_NAME, num_steps: int = 25, guidance_scale: float = 5, adapter_conditioning_scale: float = 0.8, adapter_conditioning_factor: float = 0.8, seed: int = 0, progress=gr.Progress(track_tqdm=True), ) -> PIL.Image.Image: image = image.convert("RGB") image = TF.to_tensor(image) > 0.5 image = TF.to_pil_image(image.to(torch.float32)) prompt, negative_prompt = apply_style(style_name, prompt, negative_prompt) generator = torch.Generator(device=device).manual_seed(seed) out = pipe( prompt=prompt, negative_prompt=negative_prompt, image=image, num_inference_steps=num_steps, generator=generator, guidance_scale=guidance_scale, adapter_conditioning_scale=adapter_conditioning_scale, adapter_conditioning_factor=adapter_conditioning_factor, ).images[0] return out gr.Markdown("# Doodly - T2I-Adapter-SDXL **Sketch**") with gr.Row(): with gr.Column(): with gr.Group(): image = gr.Image( type="pil", image_mode="L", shape=(1024, 1024), brush_radius=4, height=440 ) prompt = gr.Textbox(label="Prompt") style = gr.Dropdown(label="Style", choices=STYLE_NAMES, value=DEFAULT_STYLE_NAME) run_button = gr.Button("Run") with gr.Accordion("Advanced options", open=False): negative_prompt = gr.Textbox( label="Negative prompt", value=" extra digit, fewer digits, cropped, worst quality, low quality, glitch, deformed, mutated, ugly, disfigured", ) num_steps = gr.Slider( label="Number of steps", minimum=1, maximum=50, step=1, value=25, ) guidance_scale = gr.Slider( label="Guidance scale", minimum=0.1, maximum=10.0, step=0.1, value=5, ) adapter_conditioning_scale = gr.Slider( label="Adapter conditioning scale", minimum=0.5, maximum=1, step=0.1, value=0.8, ) adapter_conditioning_factor = gr.Slider( label="Adapter conditioning factor", info="Fraction of timesteps for which adapter should be applied", minimum=0.5, maximum=1, step=0.1, value=0.8, ) seed = gr.Slider( label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, ) randomize_seed = gr.Checkbox(label="Randomize seed", value=True) with gr.Column(): result = gr.Image(label="Result", height=400) inputs = [ image, prompt, negative_prompt, style, num_steps, guidance_scale, adapter_conditioning_scale, adapter_conditioning_factor, seed, ] prompt.submit( fn=randomize_seed_fn, inputs=[seed, randomize_seed], outputs=seed, queue=False, api_name=False, ).then( fn=run, inputs=inputs, outputs=result, api_name=False, ) negative_prompt.submit( fn=randomize_seed_fn, inputs=[seed, randomize_seed], outputs=seed, queue=False, api_name=False, ).then( fn=run, inputs=inputs, outputs=result, api_name=False, ) run_button.click( fn=randomize_seed_fn, inputs=[seed, randomize_seed], outputs=seed, queue=False, api_name=False, ).then( fn=run, inputs=inputs, outputs=result, api_name=False, ) if __name__ == "__main__": demo.queue(max_size=20).launch(auth=("gini", "pick"))